Hybrid Multi-Agent Deep Reinforcement Learning for Active-IRS-Based Rate Maximization Over 6G UAV Mobile Wireless Networks

Intelligent reflecting surface (IRS) can significantly improve unmanned aerial vehicle (UAV) communication qualities by reconfiguring or even optimizing wireless propagation environments. For active IRS (AIRS) enabled UAV-based mobile wireless networks, we aim to maximize the sum transmission rates of all IoT devices (IoTDs), by jointly optimizing the AIRSs’ phase shifts and amplification factors and UAV trajectory under the discrete IRS model. First, we formulate a sum transmission rate maximization problem for IoTDs, where all IoTDs can transmit data simultaneously by integrating the non-orthogonal multiple access (NOMA) and frequency division multiple access (FDMA) techniques. The formulated problem is non-convex because the coupled phase shifts and amplification factors are both discrete variables based on the discrete IRS model. Hence, we develop multi-agent deep reinforcement learning (MADRL) based schemes by integrating the deep deterministic policy gradient (DDPG) and Dueling Deep Q Network (Dueling DQN) techniques to derive the UAV trajectory and the amplification factors and phase shifts of AIRSs, respectively. Finally, we validate and evaluate the performances improvements of our developed MADRL-based schemes through extensive simulations.